This repository contains the solution developed for Ford's Challenge at the AI Impact Hackathon. The project focuses on predicting future vehicle Diagnostic Trouble Codes (DTCs) by analyzing historical fault sequences and vehicle telemetry.

Predicting vehicle faults before they become critical failures is a major challenge in automotive maintenance. To solve this, we built a hybrid predictive model that learns from sequential fault data to forecast the most likely trouble codes a vehicle will experience in the near future.

Our solution combines deep learning with statistical probability to maximize accuracy:

• LSTM Neural Network: A recurrent neural network that processes historical sequences of primary DTCs. It also incorporates engineered features like the time elapsed between faults (delta_t), odometer changes (delta_odo), and the global frequency of specific faults (log_freq).
• Markov Chain Blending: A Markov model calculates the historical transition probabilities from one fault to the next. These probabilities are blended directly with the LSTM's predictions (at a 40% weight) to ground the neural network's outputs in historical statistical reality.

• Smart Preprocessing: The data pipeline handles irregular time intervals. If a vehicle experiences a gap of more than 30 days between faults, the system automatically injects a [NO_FAULT] token to accurately represent periods of healthy operation.
• Sequence Windowing: The model looks at a historical sequence length of 10 steps to predict faults that will occur within a future target window of 5 steps.
• Imbalance Handling: Because some faults are incredibly rare, the model trains using a Custom Focal Binary Cross-Entropy (BCE) Loss function (Gamma = 1.0) to prevent frequent faults from dominating the learning process.

The model was evaluated using Recall@5, meaning the system is considered successful if the actual future fault appears in the model's top 5 predictions.

• Validation R@5: Peaked at ~0.8711 during training.
• Final Test Score: The fully blended LSTM + Markov model achieved a highly accurate Recall@5 of 0.9287 on a strictly held-out test set of 20 completely unseen vehicles.

• Language: Python 3
• Core Libraries: PyTorch (with CUDA support for GPU acceleration), Pandas, NumPy, Scikit-learn

1. Clone the repository and install the required dependencies:

   pip install pandas numpy torch scikit-learn openpyxl